Hi Mohit, I also think option 3 makes sense. We should strive to keep test matrix small so that we can do fast QA runs for dev patches. We can just use jdk7 to build & test Hive1 and jdk8 to build & test Hive2. However, I am not sure of explicitly dropping support altogether for jdk7 on Hive2. We should make sure that in pom.xml java source & target compatibility is still 1.7 (which already is the case currently) so that Hive2 is still compatible with jdk7. Unit tests as I said we can run on jdk8.
Thanks, Ashutosh On Fri, Apr 8, 2016 at 4:34 PM, Mohit Sabharwal <mo...@cloudera.com> wrote: > Hi folks, > > Oracle EOL'ed (ended public updates) for Java 7 in April 2015. > In order to support Java 8, we fixed quite a bunch of tests in > HIVE-8607 (*) early last year. However, since our pre-commit tests run > on JDK7 only, any JDK8 test failures are getting ignored. As a result, > the count has crept up > < > http://ec2-174-129-184-35.compute-1.amazonaws.com/jenkins/job/HIVE-TRUNK-JAVA8/ > > > back > from zero to 125. > > Some options to address this: > > (1) Run pre-commit tests on both JDK7 and JDK8, in both 1.x > and 2.x: This will further slow down the pre-commit run. > > (2) Alternate pre-commit test runs between JDK7 and JDK8 (in > both 1.x and 2.x): It's a cheap hack. More messy, though failures > won't get ignored. > > (3) Drop support for JDK7 in Hive 2.x, i.e. build and test Hive 2.x on > JDK8 only. For Hive 1.x, continue running JDK7 pre-commit tests. > The pros are: > - Simple test matrix. > - No need to slow down test run or maintain version specific golden files > (**). > - 2.x looks like the logical place to move to JDK8. > - Users transitioning to JDK8 for all other services do not have to > maintain > multiple java versions on the cluster. > > Option (3) looks most attractive to me. > > Moving to JDK8 also lines us up better for Java 9, which is on the > horizon (Oracle will end public updates for Java 8 in Sep 2017) > > Around 100 of the latest crop of failures are due to one cause > (HIVE-13409). > I can take a pass at triaging the rest if there is consensus around what > to do overall. > > Thanks, > Mohit > > (*) Most test failures are due to JDK8 using a different hash function for > HashMap compared to JDK7. This results in (mostly benign, but hard > to debug) ordering differences in q-file output related to serialization > order of map entries, numbering of stages in query plan, etc. > > (**) In some cases, hash function related ordering differences originate > in external libraries like Avro, antlr, json ,Thrift's map, etc. for which > code > changes are either more involved or led to more test failures. To address > this, > we added support for version specific golden files (HIVE-9109). Currently, > there > are ~40 golden files with ".java1.8.out" extension. >
